CAM2 Network Camera Object Detection Dataset and Analysis

• Activity Number: 175 - Contributed Poster Presentations: Section on Statistical Learning and Data Science
• Type: Contributed
• Date/Time: Monday, July 31, 2017 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Science
• Author(s): Kent Gauen* and Yuxiang Zi and John Laiman and Nirmal Asokan and Yung-Hsiang Lu
• Companies: Purdue University and Purdue University and Purdue University and Purdue University and Purdue University
• Keywords: Machine Learning ; Object Detection ; CAM2 ; Image Processing ; Dataset ; Deep Learning

Abstract

Continuous Analysis of Many CAMeras (CAM2) is a system for analyzing streaming data from over 120,000 network cameras world-wide. However, the current state of the art image processing methods do not generalize well to data from network cameras. Network cameras include images from a variety of angles, focus, and quality making the network camera data quite different from data normally used to train and test machine learning models. Common pre-requisites for the popular datasets include that the object(s) of interest in each image are large and in the foreground of the image. The data from network cameras are often not so "clean". In order to enable analysis, we create a CAM2 dataset for the object detection of people. Our project describes the CAM2 dataset, the shortcomings of applying current state-of-the-art machine learning methods for object detection on CAM2 data, prescribes various remedies to increase object detection score, and explains what is believed to be the bottle-neck in the image processing performance.